You can't give a worse experience to people who paid more for a better GPU.

This is the question I had about this. The reason this design works per the article is that the GPU memory is inaccessible to the OS, so the decrypted content cannot be stolen.

With a unified memory architecture, is the shared GPU memory inaccessible to the CPU?

Yes, the memory controller can simply deny accesses to specific areas from the CPU while still permitting them from the GPU.

With the proper MMU settings, yes, the CPU can definitely be denied access to some memory area. This is why devices like the raspberry pi have that weird boot process (the GPU boots up, then brings up the CPU), it's a direct consequence from the SoC's set-top-box lineage.

The GPU ultimately has to output unencrypted content, it will always be possible to steal unless we manage to implement drm in human eyes

human brains actually